The authors of the paper on growth mixture modelling (GMM) give a description of GMM and related techniques as applied to antisocial behaviour. They bring up the important issue of choice of model within the general framework of mixture modelling, especially the choice between latent class growth analysis (LCGA) techniques developed by Nagin and colleagues versus GMM developed by Muthen and colleagues. LCGA specifies that all individuals in a trajectory class behave the same, whereas GMM allows for within-class varin. The authors use conventional mixture modelling tools such as BIC for choosing the number of classes. It should be noted that a better alternative is the use of a bootstrapped parametric likelihood ratio test (BLRT), made possible in Mplus Version 4. The authors overlook a basic option for model choice. This relates to the reported convergence problem of the 3-class model. The model the authors use to decide on number of classes is overly flexible, allowing full across-class variation of growth factor variances and outcome residual variances, as well as across-class variation of regressions of growth factors on covariates. In this commentary, the author contends that a better approach is to first work without covariates and hold variances equal across classes, using this model to take a first stab at number of classes. Then use the Mplus graphics to plot the individual variation around the estimated, class-specific mean curves to see which classes need class-specific variation. And then add covariates. Unfortunately, the authors do not provide histograms of their outcomes. This would probably have shown that the outcomes are very non-normal with strong floor effects. For such data, it may be more appropriate to use a mixture model version of two-part growth modelling. In two-part modelling, a more thorough job is done in terms of modelling the large number of individuals at the floor value. Two-part modelling splits the outcome into a binary part (acting out or not) and a continuous part (if acting out, how much). Typically, the log of the continuous part is considered to bring in a long tail. The two parts are analysed as parallel processes, where if the binary variable is off (no acting out), the continuous part is scored as missing. Two-part GMM may have given different results for these data because the non-normality of the outcomes is handled more thoroughly.